AI Optimization Fails When Experts Add Too Much Data

In the push toward a circular economy, where the European Union aims for full material recycling by 2050, plastics present a stubborn . Only 14% of plastics are currently recycled, partly because recycled materials have unpredictable properties compared to new ones, making it hard to create reliable compounds for manufacturing. This unpredictability forces extensive and costly experiments, consuming time, resources, and skilled labor. To tackle this, researchers turned to Bayesian optimization (BO), an AI known for efficiently finding optimal solutions with minimal data, hoping to speed up the development of recycled plastic compounds. However, their initial attempts yielded a surprising result: the AI performed worse than traditional s used by experienced engineers, highlighting a broader issue in applying advanced algorithms to real-world problems.

The key finding from this study is that incorporating expert knowledge into the AI system, through additional features from technical data sheets, actually impaired its performance. The researchers aimed to optimize a plastic compound made from four ingredients: virgin polypropylene, recycled plastics, a filler, and an impact modifier. They sought to achieve specific targets for melt flow rate (MFR), Young's modulus, and impact strength—metrics critical for processability and product durability. While BO is designed to be sample-efficient, the addition of expert-derived features expanded the problem from four dimensions to eleven, creating a high-dimensional space that the AI struggled to navigate. This led to failed optimization runs where the AI could not find parameter combinations meeting all constraints, ultimately performing worse than the engineers' manual experiments, which achieved a best MFR of 6.65 g/10min after 25 trials.

Ology involved a step-by-step approach to test BO against a baseline of expert-designed experiments. Engineers conducted 25 experiments in three batches, measuring MFR, Young's modulus, and impact strength, with shown in Figure 1. For the AI implementation, the researchers first built a Gaussian process (GP) regression model using historical data and expert knowledge, incorporating features from data sheets to predict compound properties. This model, trained on 50 instances, had predictive performance illustrated in Figure 2, but it was not very exact. They then created an oracle model based on the engineers' experiments to simulate outcomes without physical testing. The BO procedure used a log noisy expected improvement acquisition function to minimize distance to the MFR target while handling constraints, but initial runs failed due to the high dimensionality and sparse data.

Analysis of , detailed in Figure 3, shows how different BO approaches performed. The first run, using the full model with expert knowledge, failed completely because the AI could not find feasible regions. Subsequent runs, which relaxed constraints or reformulated the problem, also underperformed, with only one experiment meeting impact strength constraints and MFR values around 5.60-5.90 g/10min. The breakthrough came when the researchers simplified the model by removing expert-derived features, reducing it to the four core composition parameters. This simplified approach, using only the 25 real experiments for training, achieved a leave-one-out cross-validation error of 4.13 g/10min for MFR, 215 MPa for Young's modulus, and 2.35 kJ/m2 for impact strength. In the final run, BO outperformed the engineers, finding 10 experiments that met all constraints and achieving a best MFR of 6.13 g/10min, close to the manual result.

Of this study are significant for industries relying on AI for optimization, such as materials science and manufacturing. It demonstrates that more data or expert input is not always beneficial; in this case, it complicated the problem due to the curse of dimensionality and boundary issues in high-dimensional spaces. For regular readers, this means that AI tools must be carefully tailored to avoid overcomplication, especially in resource-limited settings. suggest a need for systematic guidelines on when to incorporate expert knowledge, potentially through diagnostic tools or structured questionnaires, to prevent similar pitfalls in other applications like drug or energy systems.

Limitations of the study, as noted in the paper, include the small dataset of only 25 to 75 experiments, which made reliable modeling challenging. The oracle model struggled with extrapolation, particularly for predicting higher impact strength values, and the boundary issue in BO led to disproportionate sampling at parameter space edges. Future work should focus on developing practical frameworks to assess BO suitability in industrial contexts, emphasizing real-world s over theoretical advancements. This research underscores that while AI holds promise for accelerating innovation, its success depends on balancing complexity with simplicity to avoid undermining its own efficiency.